How do I prevent sensitive company data from leaking through AI?

Summary

• Govern once, enforce everywhere. Databricks prevents data leakage by pairing Unity Catalog governance with the Unity AI Gateway, so the same access controls that protect your data also govern every AI interaction.
• Built-in guardrails. The gateway can detect personally identifiable information (PII) such as emails, phone numbers, and financial identifiers in requests and responses, and either block or mask it before it reaches a model. Content safety filters screen for toxic or harmful content.
• Access control that follows the user. Agents run with the requesting user's permissions, inheriting Unity Catalog row filters and column masks so they never reach data beyond that user's authorization.
• Rate limiting, payload logging, and audit. The gateway enforces usage limits and logs requests and responses to Unity Catalog tables, giving you a complete, auditable trail.
• Real-time policy enforcement. Service policies (Beta) evaluate each agent interaction and return allow, deny, or ask-for-approval verdicts.

How do I prevent sensitive company data from leaking through AI?

Sensitive data can leak through AI in several ways: prompts that send confidential records to an external model, agents that retrieve data a user should not see, or responses that expose PII. Databricks addresses these risks with a layered approach that combines Unity Catalog governance with the Unity AI Gateway, so you define policy once and enforce it consistently across data and AI. This means you do not maintain a separate policy engine for AI: the controls that protect your tables also protect your models, agents, and tools.

Why Databricks Unity AI Gateway for preventing data leakage

• PII detection with block or mask. Built-in guardrails detect personally identifiable information such as email addresses, phone numbers, and financial identifiers in both requests and responses. You choose whether to block the interaction or mask the sensitive values before they reach an external model, so confidential data does not leave your governed boundary.
• Content safety filtering. Safety guardrails screen model inputs and outputs for toxic, harmful, or inappropriate content, keeping unsafe content out of AI-powered workflows.
• Access control that follows the user. When an agent accesses internal systems, it executes with the requesting user's exact permissions rather than a shared service account. Unity Catalog row filters and column masks apply automatically, so an agent acting on a user's behalf can never retrieve data beyond that user's authorization scope.
• Rate limiting and traffic governance. The Unity AI Gateway enforces rate limits at the user and endpoint level, curbing excessive consumption that could enable bulk extraction of data.
• Payload logging and audit trail. All interactions can be logged to Unity Catalog Delta tables, capturing request and response content so you can trace behavior, investigate incidents, and demonstrate control with a complete audit trail.
• Service policies for real-time enforcement. Service policies evaluate each interaction between an agent and a model or tool in real time, both before the call and on the result, returning an allow, deny, or ask-for-human-approval verdict. Policies can be expressed as deterministic rules or as an LLM-as-a-judge, and they follow a fail-closed default so a policy that cannot evaluate correctly blocks the interaction rather than letting it through.
• Unified governance foundation. Because Unity Catalog governs data and AI assets together, attribute-based access controls, grants, and deny policies extend from your data estate to your AI applications, with consistent lineage and auditability.

Service policies and some judge-based guardrails are in Beta; capabilities may change. The Unity AI Gateway, rate limiting, payload logging, and Unity Catalog access controls are generally available.

Getting started

• Read The Unity AI Gateway is now generally available for an overview.
• See AI Gateway: the governance layer for agentic AI for how policies govern agents.
• Configure guardrails in Configure Unity AI Gateway and Govern the content of interactions.
• Set up policy enforcement with Service policies for AI securables and How AI governance works on Databricks.

FAQs

Can Databricks stop PII from being sent to an external model?

Yes. The Unity AI Gateway can detect PII such as emails, phone numbers, and financial identifiers in requests and responses, and you can configure it to either block the interaction or mask the sensitive values before they reach the model.

How does Databricks stop an AI agent from accessing data a user should not see?

Agents execute with the requesting user's permissions and inherit Unity Catalog row filters and column masks, so an agent can never retrieve data beyond that user's authorization scope.

Can I audit what AI applications send and receive?

Yes. Payload logging captures request and response content to Unity Catalog Delta tables, giving you an auditable trail to trace behavior and investigate incidents.

What are service policies?

Service policies (Beta) evaluate each agent interaction with a model or tool in real time and return an allow, deny, or ask-for-approval verdict, with a fail-closed default that blocks when a policy cannot evaluate.